Agentic AI for Genomics

Haohan Wang's group at the University of Illinois Urbana-Champaign builds agentic AI systems that act as virtual bioinformaticians. These are teams of large language model (LLM) agents that write and run analysis code on raw genomic data, check their own work, and report gene–trait associations. The goal is to make rigorous transcriptomic analysis fast enough to run across thousands of diseases at once, while keeping the precision that scientific conclusions require.

Key questions

What is a "virtual bioinformatician"?

A virtual bioinformatician is a team of LLM-based agents that carries out a bioinformatics analysis end to end: selecting datasets, preprocessing raw gene expression files, running statistical analyses, and identifying genes associated with a trait. Haohan Wang's group develops such systems, including GenoAgent and GenoMAS.

Can LLM agents analyze gene expression data reliably?

Partly, and it can be measured. The group's GenoTEX benchmark provides expert-curated analyses from bioinformaticians for evaluating agents on gene–trait association problems. On GenoTEX, the GenoMAS multi-agent system reaches a Composite Similarity Correlation of 89.13% for data preprocessing and an F1 of 60.48% for gene identification, surpassing the best prior methods by 10.61% and 16.85%.

What can agentic AI discover at scale?

Using GenoMAS, the group analyzed over 1,300 disease–condition pairs and built a pathway-based disease similarity network. It recovers known comorbidities, reveals previously undocumented cross-category links, and points to drug-repurposing opportunities for rare conditions based on molecular proximity to better-characterized diseases.

Selected projects

GenoMAS: A Multi-Agent Framework for Scientific Discovery via Code-Driven Gene Expression Analysis

Liu H, Li Y, Wang H · arXiv preprint, 2025

A team of six specialized LLM agents that combines the reliability of structured workflows with the adaptability of autonomous agents. Programming agents expand high-level guidelines into executable steps and decide at each step whether to advance, revise, bypass, or backtrack.

Composite Similarity Correlation of 89.13% for preprocessing and F1 of 60.48% for gene identification on GenoTEX
Outperforms the best prior methods by 10.61% and 16.85%, respectively
Surfaces gene–phenotype associations corroborated by the literature while adjusting for latent confounders

[Paper] [Code]

GenoTEX: An LLM Agent Benchmark for Automated Gene Expression Data Analysis

Liu H, Chen S, Zhang Y, Wang H · Machine Learning in Computational Biology (MLCB), 2025

A benchmark for automated gene expression analysis covering dataset selection, preprocessing, and statistical analysis, following computational genomics standards, with expert-curated annotations from bioinformaticians. It introduces GenoAgent, a team of LLM agents with a multi-step programming workflow and self-correction, as a baseline.

Provides analysis code and results for a wide range of gene–trait association problems
Error analysis identifies where LLM agents still fall short in genomic analysis

[Paper] [Benchmark & code] [Project page]

Discovery of Disease Relationships via Transcriptomic Signature Analysis Powered by Agentic AI

Chen K, Wang H · Pacific Symposium on Biocomputing (PSB), 2026

Uses GenoMAS to analyze transcriptomic signatures across diseases and introduces a pathway-based similarity framework that quantifies functional convergence between diseases beyond symptom-based taxonomies.

Over 1,300 disease–condition pairs analyzed automatically
Reveals known comorbidities and previously undocumented cross-category links
Shows how background conditions such as obesity and hypertension modulate transcriptomic similarity, and suggests repurposing opportunities for conditions such as autism spectrum disorder

[Paper] [Results]

SCTA: An Agentic Framework for Stable and Interpretable Target Gene Discovery from Single-Cell RNA Sequencing

Chen S, Zhu C, Zhang Y, Li Y, Xie Q, Wang H · BioKDD Workshop at ACM SIGKDD, 2026

A decision-centric agentic framework that splits single-cell target discovery into specialized agents aligned with the key decision points of the scRNA-seq pipeline, and constrains downstream reasoning with structured biological evidence.

In a study on hereditary chronic pancreatitis, full evidence integration gives the most stable target selection across independent runs among tested configurations
Recovers disease-relevant mechanisms validated in prior studies

[Paper]

Large Language Model-based Data Science Agent: A Survey

Chen K, Wang P, Yu Y, Zhan X, Wang H · Transactions on Machine Learning Research (TMLR), 2026

A survey of LLM-based agents for data science, connecting general agent design principles (roles, execution, knowledge, reflection) with practical data science workflows (preprocessing, model development, evaluation, visualization).

Dual-perspective framework linking agent design to data science processes